Study Of ECG Biometrics Based On Subspace Learning

Biometrics play an important role in security applications and have been widely used throughout the world in the past few decades.At present,biometrics,commonly used in practical applications,include face recognition,fingerprint recognition,and iris recognition.However,these biometrics can neither effectively prevent theft,nor have sufficient capabilities to prevent counterfeitingIn recent years,ECG biometrics have caught more and more attention from the industry and academia.There are two main reasons for this.Firstly,ECG signals have basic biometrics attributes,such as universality,specificity,permanence,and collectability.Secondly,it is not easy to illegally use fake and copied ECG and ECG signals have high anti-counterfeiting capabilities.With the increasing popularity of ECG devices based on finger acquisition,the acquisition of ECG signals is becoming more and more convenient.Therefore,ECG biometrics will be a promising research areaWith the development of technology,machine learning is becoming more and more widely used.The traditional ECG biometrics method based on the basic point and frequency domain features has poor recognition performance when the ECG signal quality is low,and cannot meet the needs of practical applications.To address these issues,this thesis takes subspace learning as the main technical method,and proposes two ECG biometrics methods,which effectively improve the recognition performance(1)ECG biometrics method based on subspace fusion.The ECG signal features can be divided into local features and global features.The global features mainly describe the overall properties of the ECG,and the local features mainly describe the changes in the details of the ECG.After extracting the two features,the local and global features of the ECG signal are fused by the canonical correlation analysis(CCA)method.The use of the complementary relationship between global features and local features makes it more stable to local changes such as sudden changes in ECG signals and external interference.The voting decision mechanism is adopted so that local changes are controlled in individual areas,which improves the ECG signal recognition performance and robustness.It was found through experiments that this method can achieve better ECG biometrics.(2)The learning method based on discriminative representation of multi-scale PDV features is the first attempt to convert a two-dimensional PDV vector into a multi-scale PDV feature vector and apply it to ECG biometrics.Utilizing the extracted features improves the accuracy of the recognition rate and reduces the calculation time.Through the learned objective function,multi-scale PDV features can be projected into a low-dimensional space and the discriminative information of the ECG signals can be captured.Considering using a bag-of-words model(BoW)to represent the low-dimensional features of each heartbeat and visualize it in the form of a histogram.Through experiments on a public database,the experimental results show that this method can extract features that represent the essence of identity and improve the performance of ECG biometrics.In order to realize the practical application of ECG signals in daily life biometrics,we developed the biometrics prototype system based on ECG and combined with the proposed ECG biometrics method based on discriminative representation of multi-scale PDV features.The front-end program is responsible for user interaction,and the back-end program is responsible for feature extraction of ECG signals,ECG signal storage,ECG signal pre-processing,feature learning,ECG biometrics,user logic processing,etc.It has been laid a solid foundation for future practical applications.